To extract the vessel lumen area of anastomotic structures the vessel lumen has to be segmented within the EUS sequences. Different segmentation ap-proaches can be used to extract the vessel lumen.
Zahalka et al. [65] and Ukwatta et al. [66] has proposed using intensity based active contour frameworks for segmentation of carotid arteries in cross sectional ultrasound images. Active contours use internal and external en-ergies to deform a contour to an object by minimizing the energy of the
tour. The external energies attract the contour to features of interest within the image e.g. based on gradient or regional information. The behavior of the active contour is controlled by weighing the different components in the energy formulation. Both these algorithms were developed for segmentation of 3-d image data sets. This requires that the image data is first recorded and then analyzed in an offline setting. However, EUS sequences have to be segmented for online evaluation of the stenotic rates during the primary surgery. Additionally, the algorithms do not utilize a priori knowledge of shape or appearance which may increase robustness in segmenting objects with uncertain object boundaries as in anastomotic structures.
Guerrero et al. [67] proposed using a modified Star-Kalman filter to seg-ment and track vessel structures through ultrasound sequences. The Kalman filter was used to predict ellipse parameters in subsequent frames to segment the vessel structures. However, ellipse parameters are not sufficient to de-scribe the shape variation of surgically manipulated vessel structures.
Model based approaches such as active shape models (ASM) [68] has been proposed for segmentation of structures with a characteristic appearance and shape. ASM has previously been used in other ultrasound applications e.g.
for real-time tracking of the myocardium [69] and segmentation of the com-mon carotid artery [70]. ASM uses statistical information from shape and appearance variations of objects using principal component analysis (PCA) based on training data from manual segmentations. The trained models are then used for segmentation of new images. The use of a priori statis-tical information from previous segmentations may make segmentation of anisotropic tissue information more robust while also constraining segmen-tations to known shapes. The robustness of ASM may be further improved using 2d-PCA analysis of neighboring landmarks to obtain a better global fit of the contour. [71]
Only Guerrero et al. [67] and Hansegaard et al. [69] accounted for ob-ject movement during segmentation in the mentioned algorithms. Both ap-proaches used Kalman filters to predict object movement in subsequent im-ages. However, the sudden movement of the vessel structures in between frames when the heart is beating can make movement of the vessel struc-tures difficult to predict. Instead registration-based approaches using simi-larity metrics of the intensity information from previous frames may be used to account for inter-frame vessel movement. [72]
20
5. Background Summary and Thesis Objective
5 Background Summary and Thesis Objective
Intraoperative anastomosis quality assessment can confirm patency and dis-close technical errors in CABG anastomoses. This can enable anastomosis revision during the primary surgery to improve long term anastomosis pa-tency, clinical outcome for the patients, and the cost effectiveness of CABG.
Coronary angiography can be used to determine the stenotic rates of anastomoses and is currently considered the gold standard for intraopera-tive anastomosis quality assessment. However, it is invasive, not available in most operating rooms, and the stenosis assessment is subject to signifi-cant inter-user variability and is only based on evaluating diameter percent stenosis from 2-d images of the anastomotic structures. IFI also visualizes the anastomotic structures and it is simple to perform. However, IFI only provides a semi-quantitative assessment of the anastomotic quality and it is less sensitive than coronary angiography. TTFM provides a functional as-sessment of the anastomotic quality by measuring the blood flow through the anastomosis. It is currently recommended as the clinical standard for intraoperative anastomosis quality assessment. Still, multiple factors affect the flow measurement and the interpretation depends of the experience of the user. Additionally, it can only reliably detect diameter stenoses above 75% and it cannot define the degree and location of an obstruction within an anastomosis.
EUS can visualize the anastomotic structures and allows differentiating between gross anatomic construction errors, intimal flaps, thrombus, and spasms. It has been demonstrated to have a high accuracy in detecting tech-nical errors when compared to coronary angiography. EUS images can be used to extract the stenotic rates within anastomoses to obtain an objective measurement of the anastomotic quality and predict the long term patency of the anastomosis. EUS also allows to evaluate area percent stenosis, which may provide a more accurate evaluation of the stenotic rates e.g. in anasto-motic vessel structures with an asymmetric shape. Additionally, in vivo EUS sequences can be obtained on the beating heart to evaluate real-time dynamic information of the anastomotic structures. Currently, stenotic rates have to be quantified manually from EUS sequences as no objective methods are avail-able. Analysis of EUS sequences can be time consuming and surgeons has to be trained in interpreting EUS images or rely on peer reviews from a ra-diologist post-surgery, which may limit the use of EUS in clinical practice.
Additionally, to evaluate the maximum flow capability of the anastomosis the dimensions of the anastomotic structures has to be evaluated when they are maximally distended and dynamic information of the anastomotic
struc-tures can be lost during off-line analysis of the EUS sequences. Automatic evaluation of stenotic rates in anastomoses from EUS sequences may be ob-tained by developing automatic image analysis methods to locate and extract the vessel lumen area in the anastomotic structures.
The aim of this thesis was to develop and test automatic image analysis methods to enable quantification of stenotic rates of CABG anastomoses from in vivo EUS sequences.
An automatic anastomosis segmentation algorithm has to detect vessel structures within EUS images and extract the vessel lumen area of the anas-tomotic structures during EUS sequences without user interaction. It also has to be robust in extracting anatomical structures which is affected by noise, ar-tifacts, has an anisotropic intensity appearance, missing tissue information, variations in shape and size. It also has to track vessel structures which are subject to inter-frame movement during the EUS sequences.
The thesis will be focused on extracting the vessel lumen area within the heel and toe section of anastomoses. These sites can be used to assess the blood flow to the myocardium in both end-to-side and side-to-side anasto-moses. For the study EUS sequences from 16 end-to-side anastomoses ob-tained from 12 anesthetized healthy pigs that underwent CABG was avail-able.